AI in healthcare carries less risk when the agent works away from clinical data: administrative triage, equipment quotes, maintenance of the technology fleet and search across regulatory documents. When the use case requires a patient's health data, LGPD treats that data as sensitive, and the project is born with a legal basis, data minimization and human review.
The question we hear most from people running healthcare operations is not "does AI work?". It is "can I use this without exposing myself?". The answer depends less on the model and more on which data it sees. This article sorts use cases by data sensitivity and shows what changes in each band.
One caveat before we start. Cortex IQ has published two institutional healthcare cases, MaqMedical and Medtrix Hospitalar. Neither is an AI agent. They are platforms and digital presence. They belong here because they show, in detail, where the data lives and what a healthcare buyer demands before trusting a system.
Where does AI help in healthcare without touching the medical record?
Most of the work in a healthcare operation is not clinical. It is procurement, maintenance, documentation, scheduling and supplier support. That is the layer where an agent delivers value fast, on low-sensitivity data.
Equipment procurement and quotes. Medtrix Hospitalar connects hospitals, clinics and pharmacies to 7 verified suppliers, all with an active ANVISA registration, sales restricted to companies with a CNPJ, and an electronic invoice on 100% of transactions. The buyer requests a quote through the platform or over WhatsApp. An agent in that flow answers about accessory compatibility, lead time and product documentation. The data it handles belongs to companies and catalogs, not patients.
Clinical engineering. MaqMedical's Clinical Engineering division manages the hospital technology fleet in 8 units across São Paulo and Santa Catarina: inventory, preventive and corrective maintenance, calibration and audit documentation. Predicting equipment failure from maintenance history, or assembling the audit report from work orders, is model and agent work on machine data.
Regulatory content. The manager hiring clinical engineering wants to see compliance with ANVISA's RDC 509/2021, ABNT NBR 15943 and CFM resolution 1.821/2007. The health secretary evaluating a management system wants to know whether it is aligned with the co-financing ordinances. An agent that answers those questions from the company's documentation sees no personal data at all.
Administrative triage. Appointment confirmation, rescheduling, exam preparation instructions, questions about insurance coverage. Here the agent already touches name, phone number and time slot, which are personal data, but not yet a diagnosis. Sensitivity goes up one step, and so does the required care.
What does LGPD require when the data is health data?
LGPD classifies health-related data as sensitive personal data, in the same category as racial origin, religious belief and genetic data. That changes three things for an AI project.
First, the legal basis. Sensitive data cannot be processed under the generic justification of legitimate interest. The permitted grounds are narrower: specific and highlighted consent from the data subject, or the protection of health in a procedure carried out by health professionals and services, among others set out in the law. An agent that reads medical records to answer a patient must fit one of those grounds before the first line of code.
Second, purpose and minimization. Data collected to care for a patient cannot become, without a new basis, input to train a model or to segment a campaign. And the agent should only receive the minimum needed for the task. If the task is confirming an appointment, it does not need the lab report.
Third, the data subject's right regarding automated decisions. Anyone who has a decision made solely by a machine can request a review. In healthcare, this means any agent that classifies, prioritizes or denies something for a patient needs a recorded human review path.
On top of that come the obligations that apply to any personal data: security measures, records of processing operations, an impact report when the risk is high, and incident notification to the authority and to the data subject. The general principles are in the post on LGPD and generative AI. What follows is their application to healthcare.
| Use case | Data the agent sees | Sensitivity | What LGPD requires |
|---|---|---|---|
| Equipment quotes and compatibility (Medtrix: 7 suppliers, 100% of sales invoiced) | CNPJ, catalog, ANVISA registration | Low | Ordinary legal basis, such as contract or legitimate interest, and standard security |
| Maintenance and audit of the technology fleet (MaqMedical: 8 units) | Work orders, calibration, inventory | Low | Security and traceability of the agent's actions |
| Questions about standards and public tenders (RDC 509/2021, ordinances) | Public and company documentation | None | Nothing beyond keeping the source up to date |
| Appointment confirmation and rescheduling | Name, phone, date, specialty | Medium | Clear purpose, minimization, limited retention |
| Medical record summary for the physician | Diagnosis, exams, medication | High (sensitive data) | Specific legal ground, restricted access, human review, impact report |
| Symptom triage directly with the patient | Symptoms, history, priority decision | High (sensitive data and automated decision) | All of the above, plus right to review and a named clinical lead |
The order of the table is not accidental. The further down you go, the more the project becomes a governance project and the less a technology project.
How do you start an AI agent in healthcare without exposing yourself?
The sequence we use in any regulated sector starts with the data, not the model.
- Map the data before the use case. List what the agent would need to read, write and remember. Classify each item as personal, sensitive or neither.
- Design for the smallest possible data. Many healthcare cases can be solved with pseudonymized data or with the agent receiving only the field the task needs. If the national ID can be taken out of the flow, take it out.
- Put the human at the decision point. The agent prepares, summarizes and suggests. The person approves. That settles the right to review and keeps the technical lead in control.
- Log everything. Every call to the model, every piece of data sent, every answer returned. Without logs there is no way to respond to an incident or an audit.
- Validate in a closed scope. A 3 to 6 week PoC on low-sensitivity data proves the value and shows what governance is missing before moving up to clinical data.
Cortex IQ's AI agents service follows this order because the cost of getting it wrong in healthcare is not only technical. It is trust, and trust in healthcare takes years to build.
Three mistakes that show up often
Using a public model with patient data. Sending medical records to a consumer chat, with no data processing agreement and no control over retention, is the most common mistake and the easiest to avoid.
Treating consent as a formality. An "I accept the terms" at sign-up does not cover the use of health data by an agent. Consent for sensitive data must be specific to that purpose.
Skipping the health professional. An agent that answers patients about symptoms with no named physician behind it is not only an LGPD problem. It is a clinical liability problem.
Next step
If your healthcare operation has repetitive administrative, procurement or clinical engineering work, that is the place to start with AI. The healthcare page shows the projects we have already done in the sector, and the AI agents page explains how we structure an agent from data to production.